I built this because I was tired of the "AI writes code → I run and check → report back → repeat" loop in Godot game development.
Existing Godot MCP servers (~13 tools) focus on file operations — they can't even launch the game. I needed AI to handle the full build-test-fix cycle autonomously, so I built one with 84 tools including input simulation, runtime screenshots, and live state inspection.
What happens in the demo:
- One prompt describing a Reversi game with rules, rendering, UI, and "playtest it yourself"
- AI writes ~200 lines of GDScript, builds the scene, launches the game
- AI plays 4 moves by clicking cells, taking screenshots after each to verify
- On moves 3 and 4, AI clicks invalid cells — detects failure from the screenshot (score didn't change), reads the board highlights to find valid moves, self-corrects
Architecture:
AI (Claude/Cursor/etc) ←—MCP/stdio—→ Node.js server ←—WebSocket—→ Godot editor plugin
Key design decisions:
- All mutations go through Godot's EditorUndoRedoManager → full Ctrl+Z
- WebSocket with 10s heartbeat + exponential backoff reconnect
- Smart type parsing: AI sends "Vector2(100,200)" or "#ff0000" as strings, auto-converted to Godot types
- Tool responses include structured errors with suggestions so AI can self-correct
The four tool categories that close the loop:
1. play_scene / stop_scene — launch and control the game
2. simulate_mouse_click / simulate_key / simulate_sequence — send input
3. get_game_screenshot — AI sees the running game
4. get_game_scene_tree / get_game_node_properties — read runtime state
Existing Godot MCP servers (~13 tools) focus on file operations — they can't even launch the game. I needed AI to handle the full build-test-fix cycle autonomously, so I built one with 84 tools including input simulation, runtime screenshots, and live state inspection.
*Demo video* (4 min — AI builds Reversi from an empty project, then playtests it): https://youtu.be/D-jqmczINnQ
What happens in the demo: - One prompt describing a Reversi game with rules, rendering, UI, and "playtest it yourself" - AI writes ~200 lines of GDScript, builds the scene, launches the game - AI plays 4 moves by clicking cells, taking screenshots after each to verify - On moves 3 and 4, AI clicks invalid cells — detects failure from the screenshot (score didn't change), reads the board highlights to find valid moves, self-corrects
Architecture:
Key design decisions: - All mutations go through Godot's EditorUndoRedoManager → full Ctrl+Z - WebSocket with 10s heartbeat + exponential backoff reconnect - Smart type parsing: AI sends "Vector2(100,200)" or "#ff0000" as strings, auto-converted to Godot types - Tool responses include structured errors with suggestions so AI can self-correctThe four tool categories that close the loop: 1. play_scene / stop_scene — launch and control the game 2. simulate_mouse_click / simulate_key / simulate_sequence — send input 3. get_game_screenshot — AI sees the running game 4. get_game_scene_tree / get_game_node_properties — read runtime state
84 tools total across 14 categories (project, scene, node, script, editor, input sim, runtime, animation, tilemap, theme, shader, batch ops, profiling, export).
$5 one-time. Happy to answer questions about the architecture or MCP implementation details.